A multi-threshold semi-supervised road abnormal disease intelligent detection method based on Yolo V5

Abstract

The application discloses a kind of based on Yolo V5's multi-threshold semi-supervised road abnormal disease intelligent detection method, including collecting road surface abnormal disease image, composition teacher sample set, and the pre-processing of teacher sample set is carried out, obtains initial training data set;Yolo V5 target detection network is trained, and initial teacher model is obtained;Further collect real scene road surface image data, based on multi-threshold mode, form student sample set;Teacher sample set is carried out strong data enhancement, student sample set is carried out weak data enhancement, on the basis of initial teacher model, the training of model is carried out, and evolution teacher model is obtained;Real scene road surface image data is continuously collected, and evolution teacher model is constantly optimized, and finally target road surface disease recognition detection network is obtained;Road abnormal disease intelligent detection is carried out.The application can continuously realize the iteration of model at lower cost, and can effectively improve the recognition accuracy of various road surface diseases.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a multi-threshold semi-supervised intelligent detection method for road anomalies and defects based on YOLO V5. Background Technology

[0002] In recent years, routine maintenance of highway pavements has become a crucial issue in effectively extending the service life of roads and is key to extending the operational life of road infrastructure. However, routine inspections of common road surface defects on ordinary highways still rely on manual surveys. This method is highly susceptible to subjective factors in defect assessment and suffers from low efficiency. Currently, computer vision-based technologies are being developed to automate the identification of road surface defects, replacing traditional manual surveys. This significantly improves the detection efficiency and frequency of road surface problems, enabling early detection and treatment of defects and preventing severe performance degradation caused by delayed intervention.

[0003] Currently, computer vision-based algorithms mainly fall into two categories: The first category utilizes feature manipulation operators such as Cascade+Hog and DPM+Harr, along with their improvements and optimizations. These methods suffer from drawbacks including the need for manual feature design, resulting in a massive workload. Furthermore, these algorithms are highly dependent on lighting conditions, and their performance deteriorates sharply when classifying and identifying different types of pavement defects on complex roads or in situations with cluttered foregrounds and backgrounds. The second category, developed in recent years, leverages deep learning methods in object detection tasks and the release of open-source datasets on pavement cracks and defects by numerous research institutions. Large-scale convolutional neural networks pre-trained on massive datasets have extracted image features of far superior quality to traditionally hand-designed features. However, these solutions are limited by the number of samples collected. They are affected by various factors such as differences in geographical conditions of the collected samples, the complexity and variability of road surface defects in actual scenarios, and differences in perspective during actual image acquisition. As a result, the performance of convolutional neural network methods will be severely degraded on images outside the training dataset. Often, the recognition accuracy can be maintained above 90% on data with the same distribution as the training dataset, but on images outside the training dataset, the recognition accuracy will drop significantly to 50%-60%. Therefore, the false positives and false negatives caused by this seriously affect the application of convolutional neural network methods in real-world scenarios.

[0004] In addition, the data labeling of road surface defects is an expensive and slow process that requires the participation of a large number of business experts with domain knowledge. Therefore, in many scenarios, it is impossible to efficiently collect more and more realistic samples, which further limits the development of the above solutions. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the shortcomings of the existing technology by providing a multi-threshold semi-supervised intelligent detection method for road anomalies based on YOLO V5. By combining a small number of expert-annotated road surface defect datasets, a semi-supervised learning scheme is used to collect unlabeled image samples in various real-world scenarios. This overcomes the problems of model generalization ability and degraded performance in out-of-domain image detection in the existing technology. The method also enables intelligent mining of defect data from road surface image data under different road surface scenarios and geographical locations, thereby gradually improving the recognition accuracy of road anomalies.

[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0007] A multi-threshold semi-supervised intelligent detection method for road anomalies based on YOLO V5 includes:

[0008] Step 1: Collect images of road surface defects to form a teacher sample set, and preprocess the teacher sample set to obtain the initial training dataset D_T;

[0009] Step 2: Train the YOLO V5 object detection network using the initial training dataset to obtain the initial teacher model M_T;

[0010] Step 3: Further collect real-world road surface image data, and use a teacher model to perform pseudo-labeling of the images based on a multi-threshold method to form a student sample set D_S;

[0011] Step 4: Perform strong data augmentation on the teacher sample set D_T and weak data augmentation on the student sample set D_S. Train the model based on the initial teacher model M_T to obtain the evolved teacher model M_T'.

[0012] Step 5: Continuously collect real-world road surface image data, repeat steps 3-4, continuously optimize and evolve the teacher model M_T', improve the model's recognition ability, and finally obtain the target road surface defect recognition and detection network;

[0013] Step 6: Use the target pavement defect identification and detection network to perform intelligent detection of abnormal road defects.

[0014] To optimize the above technical solution, the specific measures also include:

[0015] Step 1 above includes:

[0016] Step 1.1: Collect images of abnormal road surface defects captured by high-definition imaging equipment, have industry experts annotate the abnormal defects in the images using experience, mark the location of the defects with square bounding boxes, and mark the type of defects with text labels to form a teacher sample set;

[0017] Step 1.2: Resize the images in the teacher sample set, and also resize the corresponding disease labeling labels to unify them to the same resolution size [1333, 800].

[0018] Step 1.3: Randomly select several normal road surface images without defect markings and merge them into the teacher sample set as background samples according to the ratio. The ratio of defect samples to background samples is 8:1.

[0019] Step 1.4: Perform horizontal flipping, 90-degree and 270-degree rotation, integer translation and histogram equalization on the image from Step 1.3 to obtain the augmented teacher sample set data;

[0020] Step 1.5: Normalize the augmented teacher sample set data and the teacher sample set data in Step 1.3, mapping the pixel value range from [0,255] to [0,1].

[0021] Step 1.6: Divide the teacher sample set obtained in Step 1.5 into a training set D_T_train and a test set D_T_val, where the ratio of training set data to validation set data is 9:1. The training set D_T_train and the test set D_T_val constitute the initial training dataset D_T.

[0022] Step 2 above includes:

[0023] Step 2.1: Train the YOLO V5 object detection network using the Adam optimizer, and set the maximum number of iterations N;

[0024] The network error is calculated by updating a specified number of Batch D_T_train samples in each iteration;

[0025] Step 2.2: When the current iteration number reaches a multiple of 10, first fix the parameters of the YOLO V5 object detection network and evaluate the objective performance indicators of the test set D_T_val data;

[0026] When the preset maximum number of iterations is reached or the objective performance index of the test set D_T_val data no longer improves, the network training is terminated, and the initial teacher model M_T is obtained.

[0027] The network error mentioned above is the sum of the following two parts:

[0028] 1) Bounding box regression loss L_bbox: Smooth L1 loss between the disease bounding box coordinates predicted by the network and the disease bounding box coordinates annotated by experts;

[0029] 2) Bounding box category loss L_cls: Cross-entropy loss between the disease bounding box category codes predicted by the network and the disease bounding box category codes annotated by experts.

[0030] Step 3 above includes:

[0031] Step 3.1: Collect real-world road surface image data again as the original detection image;

[0032] Step 3.2: Perform horizontal flipping, 90-degree rotation, and 270-degree rotation on the original detection image to obtain the test augmented image set. After normalizing the test augmented image dataset, a test sample set is formed.

[0033] Step 3.3: Use the initial teacher model M_T obtained in Step 2 to predict the test sample set;

[0034] Step 3.4: For the output results of the initial teacher model, first use the high score confidence threshold θ0 to screen the effective samples in the first stage, retain the output results with a score confidence higher than the threshold θ0, delete the output results with a score confidence lower than the threshold θ0, and select the images that are not empty in the retained output results as effective samples.

[0035] In step 3.5, among the valid samples obtained in step 3.4, the initial teacher model M_T trained in step 2 is used to perform pseudo-labeling with a low score confidence threshold θ1 to form a student sample set D_S.

[0036] The output of the initial teacher model described above is [confidence score, bounding box category code, X-coordinate of the top left corner of the bounding box, Y-coordinate of the top left corner of the bounding box, X-coordinate of the bottom right corner of the bounding box, Y-coordinate of the bottom right corner of the bounding box];

[0037] In step 3 above, the low-score confidence threshold θ1 = 0.2 and the high-score confidence threshold θ0 = 0.9.

[0038] Step 4 above includes:

[0039] Step 4.1: Perform strong data augmentation on the training set D_T_train to obtain D_T_train';

[0040] The strong data enhancement includes: image horizontal flipping, image 90-degree rotation, image 270-degree rotation, image random erasure, image color transformation, image histogram equalization, image geometric transformation, image random noise enhancement, and image JPEG artifact enhancement.

[0041] Step 4.2: Perform weak data augmentation on the student sample set D_S to obtain D_S';

[0042] The weak data enhancement includes: horizontal image flipping, 90-degree image rotation, and 270-degree image rotation;

[0043] Step 4.3: Using the initial teacher model M_T as the pre-trained model for the YOLO V5 object detection network, set a small learning rate, and train the model using a mixed dataset of D_T_train' and D_S' to obtain the evolved teacher model M_T'.

[0044] The above-mentioned random image erasure refers to randomly setting the pixel values ​​of some locations in the image to 0, while image color transformation refers to randomly changing the pixel values ​​of certain channels in the image.

[0045] Geometric transformation of an image refers to elastic transformation of the image, where the parameters of elastic transformation are: Gaussian filter parameters α = 120, σ = 6, and affine transformation parameter range (-3.6, 3.6).

[0046] The random noise enhancement of the image was Gaussian noise enhancement, with a variance range of [15.0, 50.0] and a mean of 0.5;

[0047] The image quality compression factor used for JPEG artifact enhancement is [15, 85].

[0048] Step 5 above includes:

[0049] Step 5.1: Continuously collect real-world road surface image data, iterate through Steps 3-4 repeatedly, and continuously optimize and evolve the teacher model M_T';

[0050] Step 5.2: Continue expert annotation and correction. When the number of newly annotated images by experts is 1 / 10 of the number of images in the initial training dataset D_T, expand the teacher dataset D_T and repeat steps 2-4 to obtain the target pavement defect identification and detection network.

[0051] The present invention has the following beneficial effects:

[0052] This invention, with a limited number of samples labeled by domain experts, mines inherent features from massive amounts of real-world data through the network's own learning. Using a semi-supervised learning approach, it improves network performance with lower labeling costs and higher learning efficiency, reducing missed and false detections of road surface defects. Specifically:

[0053] 1. Using YOLO V5 as the basic structure of the object detection network has low requirements for computing performance, can be quickly deployed on a large number of embedded development board modules, and can be quickly applied to real-world scenarios to acquire image data at a low cost. The network has high efficiency in semi-supervised learning.

[0054] 2. The semi-supervised learning method adopted in this invention fully explores the characteristics of massive unlabeled real-world image data, enabling rapid accumulation of samples and effectively improving the recognition accuracy of unseen road surface scenes and unseen road surface defects. Attached Figure Description

[0055] Figure 1 This is a system configuration diagram of the present invention;

[0056] Figure 2 This is a rendering of the teacher sample set for this invention.

[0057] Figure 3 This is a diagram illustrating the effect of the strong data augmentation of the teacher sample set in this invention;

[0058] Figure 4 This is a diagram showing the effect of weak data augmentation on a student sample set according to the present invention;

[0059] Figure 5 This is a diagram showing the actual testing results of the present invention. Detailed Implementation

[0060] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] To address the shortcomings of existing deep learning methods in real-world scenarios, such as the high cost and scarcity of expert annotations and poor generalization of data from other domains, this application provides a model training method that mines inherent features from massive amounts of real-world data through the network's own learning, even with only a small number of domain expert-labeled samples. This method can improve network performance and reduce missed and false detections of road surface defects by achieving lower annotation costs and higher learning efficiency.

[0062] A multi-threshold semi-supervised intelligent detection method for road anomalies based on Yolo V5 is applied to an electronic device. In this embodiment, the electronic device can be understood as a desktop computer, a laptop computer, or a high-performance GPU server, but is not limited to these types of electronic devices.

[0063] The multi-threshold semi-supervised intelligent detection method for road anomalies based on YOLO V5 can be applied, but is not limited to, using only the YOLO V5 model. Various common object detection methods can be trained using the training method proposed in this case study. The process of this invention is as follows: Figure 1 As shown. A multi-threshold semi-supervised intelligent detection method for road anomalies based on YOLO V5 includes:

[0064] Step 1: Collect images of road surface defects to form a teacher sample set, and preprocess the teacher sample set to obtain the initial training dataset D_T;

[0065] Step 2: Train the YOLO V5 object detection network using the initial training dataset to obtain the initial teacher model M_T;

[0066] Step 3: Further collect real-world road surface image data, and use a teacher model to perform pseudo-labeling of the images based on a multi-threshold method to form a student sample set D_S;

[0067] Step 4: Perform strong data augmentation on the teacher sample set D_T and weak data augmentation on the student sample set D_S. Train the model based on the initial teacher model M_T to obtain the evolved teacher model M_T'.

[0068] Step 5: Continuously collect real-world road surface image data, repeat steps 3-4, continuously optimize and evolve the teacher model M_T', improve the model's recognition ability, and finally obtain the target road surface defect recognition and detection network;

[0069] Step 6: Use the target pavement defect identification and detection network to perform intelligent detection of abnormal road defects.

[0070] In this embodiment, step 1 includes:

[0071] Step 1.1: Collect images of road surface anomalies from high-definition cameras. Industry experts will annotate the anomalies in the images using their experience. Square bounding boxes will be used to mark the locations of the anomalies, and text labels will be used to indicate the type of anomaly, such as... Figure 2 As shown, this constitutes the teacher sample set;

[0072] Step 1.2: Resize the images in the teacher sample set, and also resize the corresponding disease labeling labels to unify them to the same resolution size [1333, 800].

[0073] Step 1.3: Randomly select several normal road surface images without defect markings and merge them into the teacher sample set as background samples according to the ratio. The ratio of defect samples to background samples is 8:1.

[0074] Step 1.4: Perform horizontal flipping, 90-degree and 270-degree rotation, integer translation and histogram equalization on the image from Step 1.3 to obtain the augmented teacher sample set data;

[0075] Step 1.5: Normalize the augmented teacher sample set data and the teacher sample set data in Step 1.3, mapping the pixel value range from [0,255] to [0,1].

[0076] Step 1.6: Divide the teacher sample set obtained in Step 1.5 into a training set D_T_train and a test set D_T_val, where the ratio of training set data to validation set data is 9:1. The training set D_T_train and the test set D_T_val constitute the initial training dataset D_T.

[0077] In this embodiment, step 2 includes:

[0078] Step 2.1: Train the YOLO V5 object detection network using the Adam optimizer, and set the maximum number of iterations N;

[0079] The network error is calculated by updating a specified number of Batch D_T_train samples in each iteration. In this example, the Batch is set to 64, meaning the network error is updated every 64 D_T_train samples.

[0080] Step 2.2: When the current iteration number reaches a multiple of 10, first fix the parameters of the YOLO V5 object detection network and evaluate the objective performance indicators of the test set D_T_val data;

[0081] When the preset maximum number of iterations is reached or the objective performance index of the test set D_T_val data no longer improves, the network training is terminated, and the initial teacher model M_T is obtained.

[0082] In this embodiment, the network error is the sum of the following two parts:

[0083] 1) Bounding box regression loss L_bbox: Smooth L1 loss between the disease bounding box coordinates predicted by the network and the disease bounding box coordinates annotated by experts;

[0084] 2) Bounding box category loss L_cls: Cross-entropy loss between the disease bounding box category codes predicted by the network and the disease bounding box category codes annotated by experts.

[0085] The network error mentioned above:

[0086] L = L_bbox + L_cls

[0087]

[0088] L cls = -y*log(p) - (1-y)*log(1-p)

[0089] Where X_det and Y_det represent the horizontal and vertical coordinates of the automatically identified bounding box, respectively; x represents the difference in bounding box coordinates; y represents the label of the sample, 1 for positive class and 0 for negative class; and p represents the probability that the sample is predicted as positive.

[0090] In this embodiment, step 3 includes:

[0091] Step 3.1: Collect real-world road surface image data again as the original detection image;

[0092] Step 3.2: Perform horizontal flipping, 90-degree rotation, and 270-degree rotation on the original detection image to obtain the test augmented image set. After normalizing the test augmented image dataset, a test sample set is formed.

[0093] Step 3.3: Use the initial teacher model M_T obtained in Step 2 to predict the test sample set;

[0094] Step 3.4: For the output results of the initial teacher model, first use the high score confidence threshold θ0 to screen the effective samples in the first stage, retain the output results with a score confidence higher than the threshold θ0, delete the output results with a score confidence lower than the threshold θ0, and select the images that are not empty in the retained output results as effective samples.

[0095] In step 3.5, among the valid samples obtained in step 3.4, the initial teacher model M_T trained in step 2 is used to perform pseudo-labeling with a low score confidence threshold θ1 to form a student sample set D_S.

[0096] In the embodiment, the output of the initial teacher model is [confidence score, bounding box category code, X-value of the top left corner of the bounding box, Y-value of the top left corner of the bounding box, X-value of the bottom right corner of the bounding box, Y-value of the bottom right corner of the bounding box];

[0097] The low-score confidence threshold θ1 = 0.2 was used, and the high-score confidence threshold θ0 was 0.9.

[0098] In this embodiment, step 4 includes:

[0099] Step 4.1: Perform strong data augmentation on the training set D_T_train to obtain D_T_train';

[0100] The strong data augmentation includes operations such as image horizontal flipping, image 90-degree rotation, image 270-degree rotation, image random erasure, image color transformation, CLAHE, image histogram equalization, image geometric transformation, image random noise enhancement, and image JPEG artifact enhancement. The effect of strong data augmentation on a teacher sample set is shown in the image below. Figure 3 As shown;

[0101] Step 4.2: Perform weak data augmentation on the student sample set D_S to obtain D_S';

[0102] The weak data augmentation includes: horizontal image flipping, 90-degree image rotation, and 270-degree image rotation. The effect of weak data augmentation on the student sample set is shown in the image below. Figure 4 As shown;

[0103] Step 4.3: Using the initial teacher model M_T as the pre-trained model for the YOLO V5 object detection network, set a small learning rate, and train the model using a mixed dataset of D_T_train' and D_S' to obtain the evolved teacher model M_T'.

[0104] In the embodiments, random image erasure refers to randomly setting the pixel values ​​of some locations in the image to 0, and image color transformation refers to randomly changing the pixel value of some channels in the image.

[0105] Geometric transformation of an image refers to elastic transformation of the image, where the parameters of elastic transformation are: Gaussian filter parameters α = 120, σ = 6, and affine transformation parameter range (-3.6, 3.6).

[0106] The random noise enhancement of the image was Gaussian noise enhancement, with a variance range of [15.0, 50.0] and a mean of 0.5;

[0107] The image quality compression factor used for JPEG artifact enhancement is [15, 85].

[0108] In this embodiment, step 5 includes:

[0109] Step 5.1: Continuously collect real-world road surface image data, iterate through Steps 3-4 repeatedly, and continuously optimize and evolve the teacher model M_T';

[0110] Step 5.2: Continue expert annotation and correction. When the number of newly annotated images by experts is 1 / 10 of the number of images in the initial training dataset D_T, expand the teacher dataset D_T and repeat steps 2-4 to obtain the target pavement defect identification and detection network.

[0111] In this embodiment, with 5,000 expert-annotated samples and 10,000 real-scene samples, the network performance improvement is shown in Table 1.

[0112] Table 1

[0113] Model mAP50 YoloV5s - Supervised Learning 0.34 YoloV5s - Multi-threshold semi-supervised learning 0.45 Performance Comparison +0.11

[0114] like Figure 5 As shown, this invention, with a small number of samples labeled by domain experts, mines the inherent features of massive amounts of real-world data through the network's own learning. By using a semi-supervised learning approach, it improves the network's performance with lower labeling costs and higher learning efficiency, thereby reducing missed and false detections of road surface defects.

[0115] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A multi-threshold semi-supervised intelligent detection method for road anomalies based on YOLO V5, characterized in that, include: Step 1: Collect images of road surface defects to form a teacher sample set, and preprocess the teacher sample set to obtain the initial training dataset D_T; Step 2: Train the YOLO V5 object detection network using the initial training dataset to obtain the initial teacher model M_T; Step 3: Further collect real-world road surface image data, and use a teacher model to perform pseudo-labeling of the images based on a multi-threshold method to form a student sample set D_S; Step 4: Perform strong data augmentation on the teacher sample set D_T and weak data augmentation on the student sample set D_S. Train the model based on the initial teacher model M_T to obtain the evolved teacher model M_T'. Step 5: Continuously collect real-world road surface image data, repeat steps 3-4, continuously optimize and evolve the teacher model M_T', improve the model's recognition ability, and finally obtain the target road surface defect recognition and detection network; Step 6: Use the target pavement defect identification and detection network to perform intelligent detection of abnormal road defects; Step 3 includes: Step 3.1: Collect real-world road surface image data again as the original detection image; Step 3.2: Perform horizontal flipping, 90-degree rotation, and 270-degree rotation on the original detection image to obtain the test augmented image set. After normalizing the test augmented image dataset, a test sample set is formed. Step 3.3: Use the initial teacher model M_T obtained in Step 2 to predict the test sample set; Step 3.4: For the output results of the initial teacher model, first use the high score confidence threshold θ0 to screen the effective samples in the first stage, retain the output results with a score confidence higher than the threshold θ0, delete the output results with a score confidence lower than the threshold θ0, and select the images that are not empty in the retained output results as effective samples. Step 3.5: In the valid samples obtained in step 3.4, the initial teacher model M_T trained in step 2 is used to perform pseudo-labeling with a low score confidence threshold θ1 to form the student sample set D_S. Step 4 includes: Step 4.1: Perform strong data augmentation on the training set D_T_train to obtain D_T_train'; The strong data enhancement includes: image horizontal flipping, image 90-degree rotation, image 270-degree rotation, image random erasure, image color transformation, image histogram equalization, image geometric transformation, image random noise enhancement, and image JPEG artifact enhancement. Step 4.2: Perform weak data augmentation on the student sample set D_S to obtain D_S'; The weak data enhancement includes: horizontal image flipping, 90-degree image rotation, and 270-degree image rotation; Step 4.3: Using the initial teacher model M_T as the pre-trained model for the YOLO V5 object detection network, set a small learning rate, and train the model using a mixed dataset of D_T_train' and D_S' to obtain the evolved teacher model M_T'.

2. The intelligent detection method for road anomalies and defects based on YOLO V5 with multi-threshold semi-supervised methods according to claim 1, characterized in that, Step 1 includes: Step 1.1: Collect images of abnormal road surface defects captured by high-definition imaging equipment, have industry experts annotate the abnormal defects in the images using experience, mark the location of the defects with square bounding boxes, and mark the type of defects with text labels to form a teacher sample set; Step 1.2: Resize the images in the teacher sample set, and also resize the corresponding disease labeling labels to unify them to the same resolution size [1333, 800]. Step 1.3: Randomly select several normal road surface images without defect markings and merge them into the teacher sample set as background samples according to the ratio. The ratio of defect samples to background samples is 8:

1. Step 1.4: Perform horizontal flipping, 90-degree and 270-degree rotation, integer translation and histogram equalization on the image from Step 1.3 to obtain the augmented teacher sample set data; Step 1.5: Normalize the augmented teacher sample set data and the teacher sample set data in Step 1.3, mapping the pixel value range from [0,255] to [0,1]. Step 1.6: Divide the teacher sample set obtained in Step 1.5 into a training set D_T_train and a test set D_T_val, where the ratio of training set data to validation set data is 9:

1. The training set D_T_train and the test set D_T_val constitute the initial training dataset D_T.

3. The intelligent detection method for road anomalies and defects based on YOLO V5 with multi-threshold semi-supervised methods according to claim 2, characterized in that, Step 2 includes: Step 2.1: Train the YOLO V5 object detection network using the Adam optimizer, and set the maximum number of iterations N; The network error is calculated by updating a specified number of Batch D_T_train samples in each iteration; Step 2.2: When the current iteration number reaches a multiple of 10, first fix the parameters of the YOLO V5 object detection network and evaluate the objective performance indicators of the test set D_T_val data; When the preset maximum number of iterations is reached or the objective performance index of the test set D_T_val data no longer improves, the network training is terminated, and the initial teacher model M_T is obtained.

4. The intelligent detection method for road anomalies and defects based on YOLO V5 using a multi-threshold semi-supervised approach, as described in claim 3, is characterized in that... The network error is the sum of the following two parts: 1) Bounding box regression loss L_bbox: Smooth L1 loss between the disease bounding box coordinates predicted by the network and the disease bounding box coordinates annotated by experts; 2) Bounding box category loss L_cls: Cross-entropy loss between the disease bounding box category codes predicted by the network and the disease bounding box category codes annotated by experts.

5. The intelligent detection method for road anomalies and defects based on YOLO V5 with multi-threshold semi-supervised methods according to claim 1, characterized in that, The output of the initial teacher model is [confidence score, bounding box category code, X-coordinate of the top left corner of the bounding box, Y-coordinate of the top left corner of the bounding box, X-coordinate of the bottom right corner of the bounding box, Y-coordinate of the bottom right corner of the bounding box].

6. The intelligent detection method for road anomalies and defects based on YOLO V5 with multi-threshold semi-supervised method according to claim 1, characterized in that, The low-score confidence threshold θ1 used in step 3 is 0.2, and the high-score confidence threshold θ0 is 0.

9.

7. The intelligent detection method for road anomalies and defects based on YOLO V5 with multi-threshold semi-supervised method according to claim 1, characterized in that, Random image erasure refers to randomly setting the pixel values ​​of some locations in an image to 0, while image color transformation refers to randomly changing the pixel values ​​of certain channels in an image. Geometric transformation of an image refers to elastic transformation of the image, where the parameters of elastic transformation are: Gaussian filter parameters α = 120, σ = 6, and affine transformation parameter range (-3.6, 3.6). The random noise enhancement of the image was Gaussian noise enhancement, with a variance range of [15.0, 50.0] and a mean of 0.5; The image quality compression factor used for JPEG artifact enhancement is [15, 85].

8. The intelligent detection method for road anomalies and defects based on YOLO V5 with multi-threshold semi-supervised method according to claim 1, characterized in that, Step 5 includes: Step 5.1: Continuously collect real-world road surface image data, iterate through Steps 3-4 repeatedly, and continuously optimize and evolve the teacher model M_T'; Step 5.2: Continue expert annotation and correction. When the number of newly annotated images by experts is 1 / 10 of the number of images in the initial training dataset D_T, expand the teacher dataset D_T and repeat steps 2-4 to obtain the target pavement defect identification and detection network.

Images